feat(phi4): optional AIE4 corelib backend for Phi-4-mini-instruct

.gitattributes

@@ -0,0 +1,52 @@
+# The AIE4 model overlays are shipped verbatim and their bytes are hashed.
+#
+# `model_list.json` records a byte size and a SHA-256 for every file under
+# `src/model_overlays/`, and the installer copies those files as they sit in
+# the checkout. Git's default `core.autocrlf=true` on Windows rewrites text
+# files on checkout, so a fresh clone gets different bytes, a different size
+# and a different hash from the ones recorded -- and the catalog then describes
+# a file that does not exist on that machine.
+#
+# Measured on the AIE4 target: `test_model_catalog` failed on
+# `record.at("size") == FileSize(path)` for the overlay files, on a clean
+# checkout of the same commit that passes on a development box whose worktree
+# happened to be LF. Nothing about the product was wrong; the checkout was.
+#
+# `-text` disables all end-of-line conversion for these paths in both
+# directions, which is the only setting that makes the recorded hash a property
+# of the commit rather than of the machine that cloned it.
+src/model_overlays/** -text
+
+# `corelib_phi4_manifest.json` is 274 KB -- 10,556 lines -- of generated
+# per-initializer records, one entry per weight tensor in the ONNX. It is
+# written by `tools/generate_phi4_corelib_manifest.py`, consumed by
+# `phi4_corelib_manifest.cpp`, and read by no human: reviewing it line by line
+# is not a thing anyone does or should do, and at that size it crowds out the
+# rest of a diff.
+#
+# `-diff` makes git treat it as binary for display only. `git diff` reports
+# that the file changed without printing its contents, and `git diff --numstat`
+# reports `-` instead of a line count. NOTHING about the file's bytes changes:
+# the size and SHA-256 recorded in `model_list.json` still describe it, and
+# `git log -p -- <path>` with `--text` still shows the contents when someone
+# genuinely needs them.
+#
+# The alternative considered and rejected was re-serialising it compactly to
+# one line. That shrinks the diff the same amount but changes every byte,
+# which invalidates the recorded hash and forces a matching edit to
+# `model_list.json` -- churn in a checked artefact, to buy a presentation fix.
+src/model_overlays/phi4-mini-it-aie4/corelib_phi4_manifest.json -text -diff
+
+# The same reasoning for the three AIE4 benchmark records. Together they are
+# ~5,900 lines of per-sample timing arrays emitted by the benchmark
+# instrument. They are NOT dead evidence -- `test_phi4_continuation_calibration`
+# reads all three, and they are the fixtures roughly fifteen of that module's
+# assertions are written against, so the calibrator is tested against the real
+# measurement rather than against invented numbers. They stay, and they stay
+# byte-for-byte, because that test is green and depends on their contents.
+#
+# But nobody reviews a timing array line by line. `-diff` keeps them out of the
+# rendered diff without touching a byte.
+docs/docs/benchmarks/phi4_aie4_baseline.json -diff
+docs/docs/benchmarks/phi4_aie4_baseline_task15_rerun.json -diff
+docs/docs/benchmarks/phi4_aie4_crossover_history.json -diff

.gitignore

@@ -1,7 +1,12 @@
-build
-build_*/
-.vscode
-__pycache__
-_site
-.jekyll-metadata
-*.csv
+build
+build_*/
+# Not root-anchored: scratch build trees get dropped next to the sources they
+# build, e.g. src/test/phi4_corelib_aie4/.build-task11/.
+.build-*/
+/src/build-*/
+.vscode
+__pycache__
+_site
+.jekyll-metadata
+*.csv
+/docs/superpowers/